resume ed 025 752 al 001 636 in this paper the author ... · ture grammar (chomsky and...
TRANSCRIPT
DOCUMENT RESUME
ED 025 752 AL 001 636
By-Reich, Peter A.The English Auxiliaries: A Relational Network Description.Yale Univ., New Haven, Conn. Linguistk Automation Project.Pvb Dare Nov 68Note-44p.EDRS Price MF -$0.25 HC-$2.30Descriptors-English, Negative Fo.-ms (Language), *Structural Analysis, *Syntax, VerbsIdentifiers-*Relational Network App. oach, Stratificational Grammar
In this paper the author attempts to show that the relational network approachis adequate to handle the relationship between interrogative and declarative wordorder, the gramatically determined "do," "not," and the order of morphemes in theverbal auxiliary. He feels that this is one step toward demonstrating that therelational network approach is a viable alternative way of formalizing our knowledgeof English grammar. (See relate" document AL 001 634.) (Author/DO)
1..r1
U.S. DEPARTMENT OF HEALTH, EDUCATION & WELFARE
OFFICE Of EDUCATION
LLJ .THIS DOCUMENT HAS BEEN REPRODUCED EXACTLY AS RECEIVED FROM THE
PERSON OR ORGANIZATION ORIGINATING IT. POINTS OF VIEW OR OPINIONS
STATED DO NOT NECESSARILY REPRESENT Of fICIAL OFFICE OF EDUCATION
POSITION OR POLICY.
THE ENGLISH AUXILIARIES:
A RELATIONAL NETWORK DESCRIPLION
Peter A. Reich
November 1968
"PERMISSION TO REPRODUCE THISMATERIAL HAS BEEN GRANTED
rBY
TO ERIC AND ORGANIZATIONS OPERATINGUNDER AGREEMENTS WITH THE U.S. OFFICE OF
EDUCATION. FURTHER REPRODUCTION OUTSIDE
THE ERIC SYSTEM REQUIRES PERMISSION OF
THEPIEINNIO OWNER."
Linguistic Automation Project
YALE UNIVERSITY
New Haven, Connecticut
AL 001 636
THE ENGLISH AUXILIARIES: A RELATIONAL NETWORK DESCRIPTION
Peter A. Reich
The paper, "Symbols, Relations, and Structural Complexity" (Reich,
1968b), discusses an alternate formalsm for context-free phrase struc-
ture grammar (Chomsky and Schutzenburger, 1963). This formalism consists
of networks of relationships, a slightly refined version of the networks
proposed by Lamb (1966a; 1966b). Three types of relations are needed
to achieve a formalism equivalent in powea- to context-free phrase struc-
ture grammars: downward concatenation (downward ordered and), downward
disjunction (downward unordered or), and upward disjunction (upward
unordered or). Two additional nodes are introduced which do not add
to the power of the system, but do contribute to simplicity of the struc-
ture and its flormal properties; namely, downward ordered and-or, and
downward optional. These networks of relationships have certain Bprmal
properties; among them are associativity, bypass representation,
commutivity, distributivity, coincidence, and reduction. The formal
properties define equivalence relationships among different networks.
Figure 1 summarizes these relationships.
All of the discussion in that paper is limited to grammatical
structure essentially equivalent to context-free phrase structure gram-
mar. Chomsky and his followers have argued, and we heartily agree, that
this is not enough. Chomsky (1957:68) writes:
SYM
BO
L
_
NA
ME
--1
A
ASS
OC
IAT
IVIT
Y
B
BY
PASS
RE
PRE
SE1,
. AT
ION
C
CO
MM
UT
IV r
r Y
D
DIS
TR
IBU
TIV
ITY
E
CO
LIC
IDE
NC
E
R
RE
DU
CT
ION
A
downward
concatenation
(downward
ordered and)
No
downward
ordered
NO
/_I
._
:17
A
downward
disjunction
(downward
unordered or)
:-
AE
AbA
la
cca t)
'
*
upward
disjunction
unordered
or)
-
15.
1)
E
ownvard
od
onaptil
- A
-1
i
=_ _
Figure 1:
Summary offormal Properties
I%)
3
We have seen that a wide variety of apparently distinct
phenomena all fall into place in a very simple and natural
way when we adopt the viewpoint of transformation analysis
and that, consequently, the grammar of English becomes
much more simple and orderly. This is the basic requirement
that any conception of linguistic structure (i.e., any model
for the form of grammars) must meet. I think that these con-
siderations give ample justification for our earlier conten-
tion that the conceptions of phrase structure are fundamentally
inadequate and that the theory of linguistic structure must be
elaborated along the lines suggested in this discussion of
transformational "analysis.
Many have concluded from this and other such statements that the
only grammars which are adequate are grammars that contain transforma-
tions. While we agree that phrase structure is fundamentally inade-
quate, we do not therefore agree that the theory of linguistic struc-
ture must be elaborated in terms of transformations. There can be no
doubt that a theory consisting of phrase structure plus transformations
is considerably superior to phrase structure alone. But, we find, 'As
the depth of the analysis increases, problems mount to the point
where they indicate a serious inadequacy in a grammar that consists
only of rewriting rules. Nor is this particular difficulty overcome,
as many others are, when we add transformational rules to the grammar.'
(Chomsky, 1965:80). It is certainly possible that phrase structure
plus transformations plus distinctive features plus projection rules
plus...will bring us closer and closer to an adequate theory. But it is
also possible that another theory could do the job easier, faster, and
S.
with considerably less theoretical structure. It is with this aim in
mind that the relational network approach is being developed.
In order for one model to be shown superior to another model, it
is necessary to show that it handles all of the data least as well as
the other theory, and some of the data considerably better, or else some
data not covered at all by the other theory. In linguistics this is an
unending task, especially since the difFerent models are continually
being modiFied. Thus in the attempt to promote our mode] here, we can
only make a small dent in the subject matter that needs to be discussed.
We choose to discuss the English 'auxiliary verbs', because Chomsky
believes that 'The study oF these auxiliary verbs turns out to be quite
crucial in the development of English grammar' (Chomsky, 1957:38). We
hope to show that we cFn obtain the same kinds of insights as can the
transformationallst.
The problem of the English auxiliaries is that such constructions
as has been taking are most efficiently described as consisting of dis-
continuous constituents. A phrase structure grammar would produce
(s) (have en) (be ing) (take), and a transformation rule would reorder
the oonstituents -- have s be en take ing. (Chomsky, 1957:38-40; Klima,
1964: 253)
In order to handle this construction, we must include some Features
in our model beyond those given in figure 1. We want to do so without
adversely affecting some of the nice features we have developed, such
as the simple complexity count and the formal operations which give us
equivalence sets. What we need is simply some new nodes. In particular,
we need upward cEljunction, diagrammed as shown in the top row oF
symbol
name
A
associativity
bypass
oommutivity
1
representation
UCJ
upward
conjunction
(upward
unordered
and)
a
distributivity
ooincidence
distribution
over
concatenation
reduction
1
UOP
upward
optional
downward
precedence
disjunction
(downward
ordered or)
EVD
free
_1-
variation
disjunction
Figure 2:
Some additional nodes needed
6
figure 2.1 What this node means is that two conditions, b and c, both
must be satisfied before a is produced. For example, both zo.and past
must be present in order to get went, and both bad and comparative must
be present in order to get worse. In terms of signals, signals must come
down both b and c, not necessarily timultaneously. A signal is sent down
a immediately after b or c comes down, whichever comes later. Feedback
which goes up a travels up b and c simultaneously. In terms of formal
properties, they are similar to concatentation, except that the relation
is oammutative. There is an additional property, distribution over oian-
catenation (see figure 2). This says that if condition d is required
before a can be realized as b followed by c, the requirement (i.e., the
upward conjunction) can be placed on the a, b, or c lines of the concat-
enation element. For the same reasons that we introduced the downward
optional element and the downward ordered and-or in Reich, 1568b, we
introduce the upward optional2and the upward unordered and-or.
(4)
re
r, 4 rer- , ,
(7-) (-31
A(Dc.7) *Dcr)t(wr)
(tAtZ)
Figure 3: Simplification leading to identification
of discontinous constituent.
7
Figure 3 shows the role of upward conjunction in the simplification
process. It is highly suggestive as a model of the process by which a
child learning his language learns to discriminate its parts, even IF
they are noncontiguous. Network (1) would represent the condition in
whiz:h the child learned that the concept of eatina conjoined with the
concept of perFect tense is expressed by haevetn, while the concept
of eating together with the concept of present tense is expressed by
We see that by applying the formal properties, we can move step by step
to the simplegtequivalent network (6). This network expresses the gen-
eralization that perfect tense is realized as the discontinuous consti-
tuent hmv n, and the concept of eating is realized in both the per-
fect and present tenses as fft. We see that the initial information we
represent in our networks is not just the form, or expression, but
also the meaning, or content. This represents a major difference betmen
our approach and that of many transformationalists. The production of
discourse consists not in moving from abstract to concrete, but rather
in a transduction process between content and expression (Gleason, 1564;
Lamb, 1966c: 562ff). In this we follow Hjelmslev (1943:47-60) and agree
with Chafe (1567: 84-89).
In order to understand the English auxiliaries, therefore, we not
only need to know the patterns of English expression, but we must have
a model of the patterns of English content. The model we shall use is
that of Halliday (1967). According to this model, there are three basic
tenses in English: past, present and future.3 Complex tenses can be
built up by ooncatenation of the simple ones. For example, future past
(as in I will have finished it ...) means that whatever event we are
talking about will in the future have taken place in the past. Halliday
8
calla this past in future. In this framework the form of the verb in
such examples as I am Helping, which as been called progressive
(e.g., Palmer, 1965:59), temporary aspect (e.g., Jocs, 1964:106), and
various other things, is present present, or present in present.
Similarly, the form or the verb in such example9 as I have died is
represented in this framework as present past, or past in present, and
I had died is represented as past past, or past in past. Halliday has
recorded examples of the concatenation of up to five tenses. For
example, He had been going to have been takin ... would be an example
of pastp_a_s;tf.tittn_-easta-eseri According to this framework, each
basic tense has two realizations. Present is realized as either
nothing (91 or be-ing. Past is realized as either ed or have-en.
Future is realized as either will or 122.119ingata. Of course, most of
these in turn have alternate realizations -- e.g., be is realized in
portmanteaus as am, is,are, etc. We shall refer to the first of the two
alternate of realizations of basic tense as type I, end the second as
type II. Type I realizations occur in simple tenses and as the first
tense in all finite constructions. Type II realizations occur in non-
finite oonstructions and in complex tenses.
We shall first consider the nonfinite constructions. These are
the only constructioas that are permitted when verbal constructions
are nominalized, as in:
To publish the journal on time-)was difficult.
Publishing the journal on time ,
The nonfinite does not include the full verb phrase. One cannot say:
*To can avoid the draft is his goal.
*Wenting to Vietnam cost him his life.
9
Halliday finds twelve tenses available to nonfinite constructions, as
shown in figure 4 (d = past, f = future, p = present, s = passive).
To each of these twelve tenses can be added a passive, which gives us
fd
dp
fp
df
dfd
fdp
dfp
dfdp
to take
to have taken
to be taking
to be going to take
to be going to have taken
to have horn taki;)g
to be going to be taking
to have been going to take
to have been going to have taken
to be going to have been taking
to have been going to be taking
to have been going to have been taking
Figure 4: The twelve nonfinite tenses
such constructions as (s) to be taken, (fds) to be going to have been
taken, (dfs) to have been going to be taken, and (dfdps) to have been
going to have been being taken. It will be granted that some of these
constructions require fairly uncommon contexts in order to be produced,
but Halliday has examples of many of these from natural conversation
recorded on tape, so we shall take his word about the data. If we dia-
gram all these and simplify, we get figure 5, which was first proposed
by Newell (1966:82).
10
Figure 5: Phrase structure component of
nonfinite construction
The finite consists of a modal auxiliary, or past, present or
futgo, concatenated with the nonfinite system. If we include this in
our diagram, we get figure 6. But the tenses in the nonfinite part
are type II, while the initial tense is type I. The fact that they have
Figure 6: Phrase structure component including
finite construction
the same meaning indicates they are brought together by a disjunction
above this point.4
We now have two independent factors controlling our
grammatical decisions whether the tense is past, present, or ftrure,
11
and whether we are in the nonfinite portion of the predicate construc-
tion or not. Thus we need upward conjunction to indicate that both fac-
tors must be combined. This is diagrammed in figure 7. We have also
included the top portion of the structure which realizes the type II
tenses. In the diagram we analyze the type II future, be going to as
(be ing) sp to, where be ino is the same construction as the type II
present. The be is the same as the be of the passive be en, and the
en of the passive is the same as the en of have en. What I mean by
this is that no matter whether the be came from be en or from be ing,
it will be realized a particular way, depending upon person, type I
tense, and number (e.g. am, is, are, etc.), but not depending on
what construction it came from. Similarly, en has many alternate
t L."
t41
es'eCt
fl
?1.(I
0 e0- 4(/ .
t.P,49"t°
Figure 7: Grammar with Type I - Type II
tense disjunction added
ell
.4NIMMNIm
12
realizations, depending on what verb it is to be associated with, but
not depending upon whether it came from be en or have en.
Let us now consider what happens when negation is present. nt is
affixed to be, have, will, or any of the modals, whichever comes first.
If none are present, a do is inserted. Thus they have gone combined
with negation yields they haven't gone, but they went yields they didn't
Esl. Figure 8 shows the grammar for figure 7 with this additional infor-
mation added. The structure we have added corresponds closely to the
way we described the facts in English. We have added concatenations
saying that after have, be, or the modals including will we insert nt
if we have a negation signal from content. In the case of 3impJe past
or present, none of these occur. In this event nt is realized following do.5
We have at this point used a type of element which we have not
previously discussed -- the downward precedence disjunction (downward
ordered or). In order to understand this we shall have to look at the
meaning of disjunction in our system more closely. When we were
dealing with context-free phrase structure grammars, we defined disjunc-
tion by saying that a signal oaming in a would go down either b or c;
the choice was a random one (Reich, 1968b: figure 8). Thus we would
generate one of the set of possible sentences at random. But of course
we don't generate sentences at random. In reality we generate (barring
errors) just that sentence which we want to generate. Our system is
selectively generative; it is generative under content control. This
notion is probably one oE the most important contributions Lamb has
made. In terms of our network structures this means that each
nonft'n'itt
iity_{ 10n
Figure 8: Grammar with nezation added
disjunction leads down to one or more upward conjunctions which determine
which one of the possibilities is realized. Thus in 9(1) for a given
segement s may signal or t may signal, but it will never happen that
both s and t signal. (The zir-gag line means there may be additional
nodes between the nodes referred to.)
114.
precedencere5tAgar.crec vetriat,'ori
Figure 9: Types of disjunction
cordct,91
This is the condition that we place on regular disjunction, but we
do not want to limit ourselves to this situation. Figure 9(2) describes
another situation. Here s or t may signal, but it is also possible that
both s and t may signal. In this situation we must define what we want
the grammar to produce. We may decide that when both signal, s is to
be realized, and t is to fail. Thus s takes precedence over t. We shall
call this precedence disjunction. An example of this is used in figure
8 at point X in the diagram. If passive (s) and negation both occur,
the passive construction is realized, and do is not needed.
In some situations we don't know what the variables are that
determine why an informant makes one choice over another. They appear
to us to be random. This is the ease of free variation. We can indi-
cate this with another disjunction node, this one to make its decision
15
at random. We shall call it the free variation disjunction, shown in
9(3). As our object is to specify as completely as possible the con-
ditions which determine precisely the expression that a particular con-
tent produces, one of our goals is to have as few free variation disjunc-
tions in our grammar as possible. However, in any practical study of
grammar, we shall have to have these around. The formal properties of
precedence and free variation disjunctions are given in figure 2.
Still another disjunction situation arises. We shall term this
the conditional disjunction, after the conditional statements in LISP
and Algol-like computer programming languages. This is the situation
in which of the two possibilities, one has a condition on it (leads to
an upward conjunction), and the other does not, as shown in 9(4).
If the condition s is satisfied, the former possibility is taken,
otherwise the latter path is chosen. This is very similar to the
precedence disjunction, although logically distinct from it. In this
paper the two situations will not be distinguished. It is possible
that certain behavioral models of behavior may require that the two
situations be kept distinct.
A similar differentiation of oancatentation into four separate
cases is also useful in defining the performance characteristics of
the relational elements. The differentiation into four cases is
based on the conditions-under which the concatenation element can fail.
The four logically distinct possibilities are pictured in figure 10.
The first possibility, 10(1)lis simply that if a signal oames in a, the
concatenation will never fail to be realized. This possibility is
common in the tactic and sign patterns of linguistic structure.
16
iricalbbk
(2)
b-ca 110114.
(3)
c--fes 11;61e,
rigure 10: Types of concatenation
b-oY- c - 10,.le.
All concatenation elements in figure 8 except the two discussed below
are of type (1). In the case of the concatenation above the modals,
labelled M in figure 81 we have an example of a type (2) concatenation.
The b wire may succeed or fail, depending upon whether or not the
higher stratum signals the presence of a modal or will, but the c wire
will always succeed.
A third possibility, 10(3)lis that the concatenation may fail as
a result of a condition on the c wire, but the b wire will always
succeed if a signal is sent down it. These are considerably rarer,
but occasionally occur in tactic patterns. An example appears in
figure 8. In the case of the concatenation element above the do
(starred in figure 8), it is the success or failure of the c wire
which determines whether the concatenation will succeed. That is,
it is the presence of negation (and other things which we will add
below) that determines the presence of do in the utterance. This,
17
then, is a type (3) concatenation element.
A fourth possibility would be that both conditions on the b wire
and conditions on the c wire could independently cause the failure of
the concatenation element. This is pictured in 10(4). I have not yet
come across this possibility in grammars attempting to describe frag-
ments of natural language.
Ideally, all four possibilities should be realized by a single
type of node. However, in the current encoding model (see appendix to
Reich, 1968a) we have not been able to accomplish this. Two different
concatenation elements have been defined, type (2) and type (3). Type (1)
can be considered as a subcase of either (2) or (3). We have arbi-
trarily considered it a case of type (2). In the figures of this
paper, type (3) concatenation is starred, otherwise we shall not be
concerned with the difference between these two types of concatenation.
One of the 'apparently distinct phenomena that Chomsky finds
falls into place is the yes-no question. What appears in the declar-
ative Bonn as Meredith has dropped acid becomes in the interrogative
Has Meredith dropped acid? When no auxiliary is present, do is added,
as it is in the case of negation. Thus the declarative Julia Child
eats TV dinners becomes Does Julia Child eat TV dinners? Chomsky
says (1957:64):
The crucial fact about the question transformation T
is that almost nothing must be added to the grammar in order
to describe it. Since both the subdivision of the sentence
that it imposes and the rule for appearance of do were required
independently for negation, we need only describe the inversion
18
effected by T in extending the grammar to account for
yes-or-no questions. Putting it differently, transforma-
tional analysis brings out the fact that negatives and
interrogatives have fundamentally the same 'structure,'and
it can make use of this fact to simplify the description of
English syntax.
We shall see that relational network analysis also brinl's out
these facts. Figure 11 contains the additional structure necessary
to account for yes-or-no questions. Let us see what we have added.
First we have added the information that a declarative consists oF
the subject followed by the finite. The only atructure necessary to
include the interrogative construction is the addition of interrocr,
a line coming from content which attaches to the finite construction,
and the addition of an and-or at Y which attaches to the subject line
at Z. The result of the former connection is that in the case oF a
yes-no interrogative the sentence begins immediately with the finite
construction. The result of the latter connection is that the subject,
if it hasn't already occurred, occurs in the same place as the nt,
and has the same effect of adding the do when other auxiliaries are
not present as does the negation. If the negation also occurs, the
order is nt followed by the subject. Thus we have added the inter-
rogative with very little additional mechanism, just as Chomsky does.
We have made two additional small changes to the grammar oF
figure 8. In figure 11 we have added the copulative, be, in the same
place as the passive (indicated by s). This has the effect of adding
Mr,
19
Figure 11: Grammar with interrogative added
the copulative be to the list of words which don't get do in interroga-
tive and negative, accounting for the fact that Chomsky is another Freud
in the negative is Chomsky isn't another Freud rather than *Chomsky
doesn't be another Freud, and similarly in the interrogative. The same
20
is not true when have is the main verb: at least in American English.
Thus Americans would say Don't you have an electric toothbrush? rather
than *Haven't you an electric toothbrush? This means that the main
verb have2
combines with the auxiliary have below point X. Both haves
become had in the past and has in the 3rd person singular present, SD
we know they come together above the structure which handles this. In
those dialects in which Haven't you an electric toothbrush? is acceptable,
the line from the main verb have would come in above X but below the
concatenation with en, as shoum in the insert of figure 11.
So far, we have developed our grammar to the point where it can
produce a good portion of what we can find in the auxiliary system.
Houmver, the segments are coming out in the wrong order. Lma_Eirla
to have been finished with my income tax by now, but.., would come out
of our grammar I d1 be ing go to have en be en finish with Chomsky
finds that transformational rules make the necessary reordering very
simple. Affix followed by verb is to be rewritten as verb followed by
affix, where verb includes modals, have, be, etc., and affix includes
past, en, ing etc. After this rule is appled there is a large number
of rules of the form: do past-odid, do en-Adan, do neg-Adönt, do sing -
ular.4daz, etc. (Chomsky, 1957: 39-40)6 We could, if we wanted to,
represent these rules in our network. However, there is a simpler way.
First of all, all of the modals and have7
, be, and do are irregular,
that is, they have alternate phonological realization depending upon
morphemic context. Why take two steps to go from past do to do past
to did when it is just as easy to state the rule past_st-Adid? In our
system this course seems simpler. And what about the generalization
21
that no matter what the morphemic context of do, the initial phoneme is
always d? The structure I propose allows one to make such generalizations.
Although such generalizatiuns are relatively trivial and uninteresting,
the formalism developed to describe linguistic structure should be
flexible enough to allow such statements. We shall see how they can
be made within the relational network framework.
Let us first look at the structure which will account for regular
verbs, shown in figure 12. We see that following the regular verb there
Figure 12: Structure for regular verbs
is the possibility of three different endings. In the environment of
past I or en, the ending is d. In the environment of present I there
are two possibilities: if additionally the environment includes aa
22
person and linp_l_AE, the ending is z; otherwise there is no ending
(or ing - we shall see how this is produced later). In the event that
neither os_t_I or en nor present I occur, the verb is realized without
an ending.
Although the diagram is a straightforward representation of the
facts, it does not fit into the patterns proposed by Lamb (1966b).
Specifically, it appears to put into one stratum information that Lamb
preferred to split into two strata, lexemic and morphemic. It is my
feeling that if the requirements for what is allowed in the realizational
portion of linguistic structure are relaxed slightly, the information
Bormerly described in these two strata can be combined into one stratum.
This is a rather involved empirical and theoretical question, the
discussion of which I shall save for a future paoer. For the purposes
of this paper the reader should simply be warned not to expect the
diagrams in this paper to conform completely with all of Lamb's 1566
hypotheses.8
Let us now tUrn to the structure of some irregular verbs, to see
how filey can be handled. Figure 13 shows the structure I propose for
do. The initial consonant is d no matter what the environment is.
There are two sets of environments to be considered. The first choice
is based on whether do occurs with present I, past I, or en. If it
is past I the vowel is i and the final consonant is d. If it is en
the vowel is a and the final consonant is n. Note that the concatena-
tion is type 3; the choice of vowel is determined by what follows. If
do occurs together with present I, then a second set of contexts deter-
mines the vowel. In the environment 3rd -person singular the vowel is a
23
and the final consonant is Z. If not (precedence disjunction), but if
the context includes Ligl (negation), the vowel is 5. Otherwise,
(conditional disjunction) the vowel is M.
Im111/0
Figure 13: Structure for do.
The structure of go and will, show in figure 14, is similar to
the structure of do. In the case of go the same three contexts deter-
mine which form is realized, but this time one form, wenD, which is
realized in the environment past I is entirely distinct from the other
two forms, which begin with R. The fact that the other two forms
differ in their vowel can be described either morphophonemically, that
is, as being determined by the environments en or present I as the case
24
Figure 14: Structure for za and will
may be, or phonemically by the presence of a following n or some such
rule. The representation shown in figure 14 is the morphophonemic one.
The D in wenD is a special morphon which devoices a following d, so
MN/Dd/ is realized as P/t/. Otherwise MNthat /D/ is realized as d.
Thus MN/benDd/ is realized as P/bent/. Similarly with send, sent, and
lend, lent. Will is realized as wd in the environment neg, otherwise
as wil.
Figure 15 gives a possible description of one of the most compli-
cated morphological structures,namely that of be. Here again we find
two levdls of choices. The first is unique in that there are four
choices. In all other verbs the infinitive andim form of the verb
is the same as that of present 1, SO that these lines come together (at
25
Figure 15: Structure for be
the upward disjunction marked X in the figure) and act as a single
conditioning environment, labelled present I in figures 12, 13, and 14.
In the verb be the forms are distinct, as dhown in figure 15. In the
second level of choices we find the familiar 3rdasim, but in addition
we also find nelzaTest (which will be discussed in detail below) and
1st-sing as conditioning environments. I will leave it to the reader
to check.out that the diagram does indded produce am, is, are, was,
were, be, being, and been when appropriate.
Thus far I have described the realizational structure of various
26
verbs as if they were separate and completely ideosyncratic. This is,
however, not the case. Go and do(and all the irregular non-modal verbs)
behave like regular verbs with respect to 3rd-sing in having the
ending z. Other facts can be generalized over fewer verbs, such as en
being realized as n in been, done, gone, taken, etc. Still other facts
apply to as few as two verbs, such as the 6 for the vowel of the verb
in the presence of neg which is only true for won't and don't. This is
no particular problem for a network grammar. The descriptions given in
figures 12 through 15 are easily combined and simplified to represent
identical structures only once, and to identify specifically the class
of items for which the structure applies. The resulting structure is
shown in figure 16. For example, the upward disjunction at X in the
center of the figure defines the class of verbs that take n as the real-
ization of en.
Again I wish to emphasize that this diagram is not necessarily
the optimal way of representing this grammatical information within a
relational framework. That is beside the point. Representing these
facts as a network seems superior to transformational grammar alterna-
tives. If one represents these facts in a lexicon; that is, a list oF
separate lexical items, one is not able to state the partial generali-
zations that occur. One is condemned to inefficiently repeat the rules
over and over again in each item for which they apply. If one writes
transformational rules to describe the partially generalizable struc-
tures, one is faced with the problem of how to specify when a transfor-
mation Is applicable. There is another problem one encounters with
this approach. Notice that while the realization of, for example, do
vey1". 1,04
0'10
ree-(see 1r`c
Figure 16: A fragment of realizational structure
is determined by the environment of en, it is also true that the reali-
zation of en is determined by do. In a rewrite rule framework, one could
28
express these facts with rules of the form:
(<3) a-tx / b and (118) b-->y / a
The problem is that these rules in a rewrite framework must be
ordered. We want ab to be realized as Isz. But if we order the rules
,(2iN'ab would be realized xb, and if we order the rules fr( ab would
be realized 22,.. One can always get around this problem in a rewrite
framework by means of a special rule of the form ab4xy, but this has
the disadvantage that within the grammar a is not related to x and b is
not related toz. But do and da are certainly related and en and n are
certainly related. Obe might propose a rule doida/ n, except that
this would result in don't being realized as dant. This can be avoided
by inserting an initial rule do->d6/ neg, so that our rules are ordered:
doodoj neg
en.in/do
do.ida/ n
neont
This has the disadvantage of separating the two rules for do. It also
forces an ordering between a rule concerning !lea and a rule concerning
en, which seems ad hoc since they never occur in the same word. It
seems to me that the best solution is to allow simultaneous application
of rules, if one must use rules at all.9
We now turn to the problem of order disparity in relational networks.
If we put together the networks of figures 11 and 16, we notice that
the original ordering of elements is not the one that is ultimately
realized. Figure 17 shows an example. We notice that the grammar above
the dashed line produces have en be ing. But the network below the
29
dashed line specifies that the order should be bi n rather than en be.
If we are not interested in a perfprmance model we can simply state
the general principle that lower level orderings take priority over
higher level orderings.
n4..
'have
Figure 17: Example of order disparity in
relational networks
How can such a principle be realized in a performance model? In
earlier papers (Reich, 1968a:2-7; Reich, 1968b: 11-14) it was explained
that concatenation ordering works by means of a feedback signal. Thus
a signal does not start down B in figure 17 until a feedback signal
30
oames up A, in effect signalling that the entire structure emanating
down from A has been successfully realized. But en can't be realized
until after a signal is sent down B. The conflict is easily resolved
by adding a node of a new type, a 'feedback reflector' at X. When
a signal comes down to a feedback reflector, it continues to send the
signal down the wire, but it also sends feedback up the wire immediately.
When feedback finally does oame up to the feedbadk reflector, it simply
stops, since feedback has already been sent up from there.
In the grammar of English auxiliaries discussed in this paper,
the feedback reflector must be added at four places, present-I,
past=1, en, and ing. This simple addition to the theory allows one to
handle both anataxis and discontinous constituents.
The last area we shall discuss is the problem of negation. The
problem arises because of the missing amn't in English. In my normal
speaking dialect I say I'm not when the subject appears before rhe
auxiliary, and aren't I in interrogatives where the pattern is -irst
auxiliary before subject. The question is wny aren't I rather than am
I not. This question leads to the problem of meaning in negation. Of
the three accounts of negation I looked at (Klimi, 1564; Palmer, 1565;
and Joos, 1964) two offered suggestions as to the meaning of not.
Palmer (1565:43) suggests that of the two forms of not, n't modifies
the modal auxiliary and not modifies the non-finite form that follows
it, and that n't should be oonsidered to precede the auxiliary to
which it is affixed. Both of these suggestions seem overly simplified.
Consider the sentences in figure 18 (the apostrophe indicates emphasis,
probably realized by intonation). We see that what is emphasized is
31
18a. Zellig didn't stay at the 'hotel.
Zellig didn't 'stay at the hotel.
18c.'Zellig didn't stay at the hotel.
18d. Zellig didn't stay 'at the hotel.
18e. Zellig 'didn't stay at the hotel.
18f. Zellig did 'not stay at the hotel.
18g. Zellig 'did not stay at the hotel.
18h. 'Didn't Zellig stay at the hotel?
18i. Did Zellig 'not stay at the hotel?
Figure 18
(He stayed at the 'inn.)
(He moved into a 'flat.)
(but 'Rulon did.)
(He stayd 'in it.)
(He commuted daily from home.)
(contrary to What you thought.)
(as you had suspected.)
(No, he didn't.)
(Yes, he did not.)
=1111S
what is negated. In 18a hotel is negated, but the rest of the sentence is
not. In 18b stay is negated. In 18c Zellig is negated. In 18d at is
negated. In 18e the entire predicate seems to be negated, and in 18f only
the polarity ,)f the sentence seems to be negated. But 18e and 18f can
both mean either; the difference seems inbe only a tendency. Thus in the
fifth previous sentence in this paragraph, I prefer is not to isn'tlbut
would accept either. In 18g not modifies stay at the hotel, but the
predicate is considered a positive act. This is best seen in 181, which
is the interrogative of 18e. In general, the parenthesized phrases or
clauses to the right of each of the examples cannot be concatenated with
examples other than the one they are with. Thus, for example, 'Zellig
didn't stay at the hotel; he moved into a 'flat. is less acceptable, if
not ungrammatical.10
Neither Palmer's nor Joos' suggestions can account for the variety
shown in examples I8a through 18g. Nor do they account for the diff-
erence between 18h and 181, since do in these sentences has no meaning.
32
Moreover, we find that not cannot occur anywhere in the verb phrase;
it can occur immediately after the first auxiliary in the form n't or
not, and it can occur immediately before the main verb.11 Thus 1,
19t, and 19d are grammatical, but 19c is not. Based on these examples
1St. She wasn't going to have typed the thesis.
19io. She was not going to have typed the thesis.
19c. *She was going to not have typed the thesis,
19d. She was going to have not typed the thesis.
Figure 19
we conclude that there are two types of negation: indefinite negation,
which is realized immediately after the first auxiliary, and which has
as its range that part of the sentence which is emphasized, and definite
negation, which indicates a positive act not to do something, and
which occurs immediately before the main verb. These two negations are
independent of one another; they can both occur in the same sentence:
Zellig didn't not stay at the hotel. When there is only one auxiliary
the two slots occur at the same place in the sentence, resulting in
ambiguity in the written, but not in the oral, forms, as in 18f versus 18g.
Given this description of not, we can understand why in the indi-
cative I say I'm not, while in the interrogative I say aren't I. The
missing amn't, corresponding to 18e is not a problem because of the
existence of the alternate form I'm not, corresponding to 18f. But
am I not doesn't work, because this is definite negation, and means
something different. Thus a contracted form must be used, and the one
33
that was dhosen by people who disdain to say ain't is aren't.
The structure Which describes the two types of negation at this
level is shown in figure 201 where Y indicates the node corresponding
4ocIAS
mk+/19
cleft
11)3
not
Styhed
4r 02.4or, nt
rtoKc e
cmos
3-0
ct.gml(
rer:i
prtference io
pm I hot
Figure 20. Structure of negation
to the node labelled Y in figure 11. Y, it will be recalled, leads to
all the features which require the realization of the empty do if modals,
have, or be are not present. Another feature which has this property
is 221arity emphasis. This also is trivially added to our grammar.
Another feature is a juncture that results after the subject if it is
needed to account for the difference in the number of syllables in 21a
r
34
versus 21b. This example was pointed out by Palmer (1965:33). It is
21a. Now I have gone. [aiv]
21b. Should I have gone [ai av)
Figure 20
an example of a place where syntax affects phonology. This is no
particular problem for the relational network approach. When we need
a syntax connected feature in phonology, we send a signal down to the
phonology from the appropriate place in syntax. There seems to be some
confusion on this point. Postal (1968:118) states: "The stratifica -
tional rejection of phonological or morphophonemic rules which refer
to Surface constituents is..among the clearest evidence of the extent
to which this theory fails to be descriptive of human language." We
do not carry into the phonology all of the information about the
'surface structure ' because we don't need all of it. But, of course,
we bring down any information that is needed. Where these lines enter
the phonology, we call them phonological features, because in general
they come from more than one place in the syntax, or go to more than one
place in phonology, or both. That is, they go through an alternation
pattern. For example, the juncture needed to account for the sentences
in figure 21 also occurs after noun phrases which are not simple pronouns,
as evidenced by the minimal pair in figure 22 where Shei is a nickname
for Sheila (Palmer, 1965:33).
35
22a. Shei'll be there. ( i: al]
22b. She'll be there. ( i: 1]
Figure 22
Postal seems to be objecting both to calling such features phonolo-
gical and to the fact that we bring down only the necessary features,
rather than all features. He feels our approach is ad hoc. However,
Chomsky (1967:108) has pointed out that evaluative notions such as
simplicity are theory specific - the same argument applies to notions
of ad hoc. From the point of view of stratificational grammar, to send
all syntactic information to the phonology when only a few junctures
are needed is like killing a mosquito by dropping an H-bomb on ft. Of
coursed in order to complete the argument, it is necessary to demon-
strate that the transformational cycle is unnecessary to account for
stress in English. This would be a paper in itself (at least), but a
gketch of the argument is given in Reich, forthcoming, and so it will
not be pursued further here.
Figure 23 gives our final tentative sketch of the grammar for
the English auxiliaries, with negation, polarity emphasis, and the
above-mentioned juncture.
The grammar described by figure 23 attached to figure 16 differs
from Newell's (1966) stratificational treatment in two ways. The first
is that it does not follow the 1566 stratificational format. This
has resulted in reducing a certain amount of redundancy that existed
in the earlier treatment. The second difference is that Newell's
treatment causes problems in defining a performance model to fit his
36
IMO
-
Figure 23: Tentative grammar for the
auxiliaries
37
formulation. The morphemic system needs to 'look ahead' to see what
is going to come down from the lexemic stratum before it actually
comes down, in order to handle the insertion of the grammatically deter-
mined do properly. This can be avoided by adding some additional
structure to the morphotactics but it increases bistratal redundancy,
as this additional structure is already in the lexotactics. In order
to demonstrate that my formulation does not suffer from such performance
problems) which can be very difficult to spot, I have tested it on the
computer, using the Relational Network Simulator (Reich, 1968).
Sample output of a slightly earlier version appears in the appendix to
that paper. The definition of the nodes used in running this network
is also given in the appendix. The network produced as required in all
samples tested.
In this paper we have shown that the relational network approach
is adequate to handle the relationship between interrogative and declar-
ative word order, the grammatically determined do, not, and the order
of morphemes in the verbal auxiliary. This is one step toward demon-
strating that the relational netuvrk approach is a viable alternative
way of formalizing our knowledge of English grammar.
Footnotes
38
1In his recent work, Lamb sometimes uses diamonds where I am going
to use upward unordered ands (upward conjunction). This is based on
two considerations. One is that if upward unordered ands are used as
I use them in this paper, they cannot be defined in the same way as
downward unordered ands. I prefer hot to be bound by a constraint
requiring that upward and downward nodes with identical logical function
need necessarily be defined identically. The other consideration is
that he is trying to solve the problem of both language production (en-
coding) and language perception (decoding). In some versions of decoding
it appears that these nodes behave differently with respect to their
two upward wires. If this is the case, then the node cannot be commu-
tative and should be represented differently. I have limited myself
in these papers to the problem of encoding, and in this domain I have
found no reason-to depart from Lamb's (19661) earlier notation.
2Both optional nodes are indicated with a circle on the optional
line. A downward optional is put near the node above it on the line,
and an upward optional is put near the node below it on the line. In
situations where this does not make the difference clear, I indicate
the upward optional by adding a line immediately below the circle, as
shown in figure 2, and the downward optional by adding a line immed-
iately above the circle, as dhown in figure 1.
31 have chosen this model merely for purposes of explicitness.
A two tense description such as that described by Huddleston, Hudson,
Winter and Henrici (1968) has much to recommend it, although I
personally don't agree with all their examples (this is probably a
39
dialect difference). David Bennett and Geoffrey Sampson, working at
Yale, have both found additional evidence for a two tense system.
Since most of the difference would appear in the networks which would
describe English semology, and since this paper is concerned exclu-
sively with lexemic and morphemic structure, itis not important to
the arguments in this paper how the issue is ultimately resolved.
411his is the point in the grammar which would have to be modified
if the semology is better described in terms of a two tense system.
Spe footnote 3.
5This design assumes the signal for negation is one of the first
signals sent down by the semology. Otherwise such non-English strings
as *I will haven't Eat instead of / won't have gone might be produced.
6This is one place in which extensive changes have taken place in
transformational theory since 1957. In my opinion these changes have
only made matters worse, at least with respect to morphophonemic
structure. I have never seen in print how transformationalists propose
to handle this structure within their current flormulation, although
Wall (personal communication) has indicated how he thinks it might be
done. In this paper I will refer only to the 1957 version of transfor-
mational grammar, and save discussion of the newer formulation and the
reason I dislike it, for a future paper.
7The irregularity of have can be handled completely in the phono-
logy by spelling it haV where V is a consonant which is realized as v
only in syllable final position. Thus MN/haVd/ is realized as 13/h ae di/
MNand /haVZ/ is realized as P/hm z/.
40
8After this paper was written Lamb modified his position so that
it might be relatively close to mine. However, it is too early to tell
exactly what the correspondence will be.
9For more discussion on this issue see Lamb, 1564:111ff, Chomsky,
1562c:97, and Lamb's description of the same data in Lamb, 1566b:39.
100ne must not confuse this example with 'Zellig didn't stay at the
hotel; 'he moved into a 'flat., which is grammatical in an appropriate
context.
11Except in the case of the cophiative be, where it occurs after
the main verb.
141
Bibliography
Chafe, W. L. (1967). Language as symbolization. Language 43:57-91.
Chomsky, N. (1957). Seltactic Structures. The Hague: Mouton.
Chomsky, N. (1962). The logical basis of linguistic theory. Paper
presented at the 9th International Congress of Linguists, Cam-
bridge, Mass., 1562. In Lunt. (Revised, expanded and retitled
'Current issues in linguistic theory' in Fodor and Katz, 50-118.
Later version with the same title published separately: 1564.)
The Hague: Mouton. Also published in 1566, New York: Humanities.
Chomsky, N. Aspects_of the Theory of Syntax. Cambridge, Mass.:
M.I.T. Press.
Chomsky, N. (1967). Some general properties of phonological rules.
Language 43:1, 102-128.
Chomsky, N. and Halle, M. (1968). The Sound Pattern of Engligh. New
York: Harper and ROW.
Chomsky, N. and Schutzenburger, M.P. (1963). The algebraic theory of
context free languages. In Braffort, P. and Hirschberg, D. (eds.).
Cs_mgj_i_gnand_FoauterProrwnrmns 118-161. Studies in
Logic Series. Amsterdam: North Holland.
Gleason, H.A. (1964). The organization of language. In Stuart, C.I.
J.M. Re ort of the Fifteenth Annual Table Meetin. on Lin istics
and Language Studies, 75-95. Washington, D.C.: Georgetown
University Press.
Halliday, M.A.K. (1967). Lectures given at Yale.
Hjelmslev, L. (1943). g1'_umg_Proleomenatoa'heoofLane. Translated
by Whitfield, F.J. Madison: University of Wisconsin Press.
42
Huddleston, R.D., Hudson, R.A., Winter, E.O. and Henrici, A. (1968).
Sentence and clause in scientific English. Report of the research
project 'The linguistic properties of scientific English', Com-
munication Research Centre, Department of General Linguistics,
University College London.
Joos, M. (1964). The English Verb. Madison: University of Wisomnsin Press.
Lamb, S.M. (1964). On alternation, transformation, realization, and
stratification. In C.1.J.M. Stuart (ed) (1964).Report of the
rifteenth Annual (Eirst International) Round Table Meeting
on Linguistics and Lanpage Studies. Monograph Series on Lan-
guages and Linguistics 17, Washington, D.C.: Georgetown University
Press, 105-122.
Lamb, S.M. (1966a). Prolegomena to a theory of language. _ALEIgRalt 42.
537-573.
Lamb, S.M. (19661). Outline of Stratificational Grammar. Washington,
D.C.: Georgetown University Press.
Lamb, S.M. (1966c). 'Epilegomena to a theory of language. Romance
Philology 19. 531-573.
Newell, L.E. (1966). Stratificational analysis of an English text.
Appendix to Lamb, S.M. (1966b). Outline of Stratificational
Grammar. Washington, D.C.: Georgetom University Press.
Palmer, R.F. (1965). A Linguistic Study of the English Verb. London:
Longmans.
Reich, P.A. (1968a). The relational network simulator. Linguistic
Automation Project Report. New Haven, Conn.: Yale University.
Reich, P.A. (1968b). Symbols, relations, and structural complexity.
Linguistic Automation Project Report. New Haven, Conn.: Yale University.
43
Reich, P.A. (forthcoming) The finiteness of language. Paper to be
presented at the 1968 LSA meeting. Linguistic Automation Project
Report. Wiw Haven, Conn.: Yale University.
Wall, R. (personal oommunication).Discussion at Communication Sciences
Seminar, University of Florida, Gainesville, September 1968.